Sports Classifier 🏅

A fastai vision learner that classifies sports photos into 100 sport classes — everything from air hockey to wingsuit flying.

This is the model behind the Game State Vision Streamlit app.

Model repo (this): mayurgp/sports-classifier
Live app: Game State Vision (Streamlit Community Cloud)
File: sports_classifier.pkl (~47 MB, fastai Learner export)

Model details

Attribute	Value
Task	Image classification (single-label, 100 classes)
Backbone	CNN (fastai `cnn_learner`, `Sequential` head)
Input size	`224 × 224`, RGB, squish-resize
Output	100 probabilities (softmax)
Format	fastai `Learner` pickle export (~47 MB)
Framework	PyTorch / fastai 2.8.8
Visualize	`mayurgp/sports-classifier`

Classes (all 100)

sports_classifier predicts exactly one of these 100 sports:

air hockey, ampute football, archery, arm wrestling, axe throwing, balance beam, barell racing, baseball, basketball, baton twirling,
bike polo, billiards, bmx, bobsled, bowling, boxing, bull riding, bungee jumping, canoe slamon, cheerleading,
chuckwagon racing, cricket, croquet, curling, disc golf, fencing, field hockey, figure skating men, figure skating pairs, figure skating women,
fly fishing, football, formula 1 racing, frisbee, gaga, giant slalom, golf, hammer throw, hang gliding, harness racing,
high jump, hockey, horse jumping, horse racing, horseshoe pitching, hurdles, hydroplane racing, ice climbing, ice yachting, jai alai,
javelin, jousting, judo, lacrosse, log rolling, luge, motorcycle racing, mushing, nascar racing, olympic wrestling,
parallel bar, pole climbing, pole dancing, pole vault, polo, pommel horse, rings, rock climbing, roller derby, rollerblade racing,
rowing, rugby, sailboat racing, shot put, shuffleboard, sidecar racing, ski jumping, sky surfing, skydiving, snow boarding,
snowmobile racing, speed skating, steer wrestling, sumo wrestling, surfing, swimming, table tennis, tennis, track bicycle, trapeze,
tug of war, ultimate, uneven bars, volleyball, water cycling, water polo, weightlifting, wheelchair basketball, wheelchair racing, wingsuit flying

Usage

1. Install dependencies

pip install fastai torch huggingface_hub pillow

2. Download & load

from huggingface_hub import hf_hub_download
from fastai.vision.all import load_learner

model_path = hf_hub_download(repo_id="mayurgp/sports-classifier", filename="sports_classifier.pkl")
learner = load_learner(model_path)

3. Predict

from fastai.vision.all import PILImage

img = PILImage.create("soccer.jpg")
pred, pred_idx, probs = learner.predict(img)

print(f"Predicted: {pred}")

# top 5 with probabilities
import pandas as pd
top = sorted(range(len(probs)), key=lambda i: probs[i], reverse=True)[:5]
for i in top:
    print(f"{learner.dls.vocab[i]:20s} {probs[i]:.2%}")

Note on Python version differences. This model was exported on Python 3.12. If load_learner fails with a plum/Resolver pickling error on a newer setup, apply the small compat shim in infer.py before loading.

Try it live

Upload a sports photo and get a prediction in the deployed Streamlit app (Game State Vision).

Security / trust note

load_learner uses Python's pickle, which can execute arbitrary code. Frameworks picklescan marks a fastai Learner pickle as Unsafe (`fastai.learner.Learner imports) — this is expected for fastai exports. Only load this file from a trusted source (this public repo).

If you need a stricter, weight-only distribution, re-export the model to safetensors / model.pth + a vocab.json, and load with learn = Learner.load(...) instead.

Dataset / reproduction

The app/repo metadata for this card: see mayurgp/game-state-vision and the streamlit-colab-app repository.
Fine-tuning with fastai's cnn_learner + Resize(224) on a 100-class sports dataset.
Export produced by learner.export() → sports_classifier.pkl.

Model card maintained at hf-model-card/README.md.

Downloads last month: -; Downloads are not tracked for this model. How to track